Spectral patterns of main classes
Due to the limited availability of meta data, we apply basic image processing strategies e.g. contrast stretching, Lillesand
and Kiefer, 1999 to the raw intensity values. For each channel, the 99.99th percentile is computed and used as intensity cutoff
threshold. This way, outliers from specular reflections are eliminated. Visual inspection of these points revealed that such
outliers occur on moving objects like cars and other man-made objects like building facades. Finally, a histogram stretch to
twice the standard deviation is computed for each channel, resulting in value ranges from 0 to 255.
Figure 1. Merged point cloud coloured as false-colour composite red=C2, green=C3 and blue=C1
The scaled intensity values are used to create false color composites for each point. The combination red=C2, green=C3
and blue=C1 is used to simulate a CIR color infrared similar visual appearance, and the combination red=C1, green=C2 and
blue=C3 for a true color similar appearance. The channels C2 and C3 are used to calculate a pseudo NDVI Eq. 1:
pseudo NDVI = 1
where pseudo NDVI =
pseudo normalized difference vegetation index C2 = near infrared channel 1,064 nm
C3 = green channel 532 nm Finally, the point cloud is classified into the main classes
‘unsealed ground’,
‘sealed ground’, ‘buildings’,
‘mid vegetation’, ‘high vegetation’, ‘water surface’ and ‘water body
bottom’ and for some small training areas into the subclasses ‘green grass’, ‘dried up grass’, ’sand bare soil’, ’wetlands’,
‘darker asphalt’ and ‘lighter asphalt’. The classification is done by a two stage approach consisting of automatic and semi-
automatic classification. First the point cloud is automatically classified into ‘ground’, ‘buildings’, and ‘mid vegetation’ and
‘high vegetation’ classes by a “geometrical classification” using information about x,y,z coordinates and neighborhood
exclusively. Ground points are classified using a hybrid approach of progressive TIN densification Axelsson, 2000 and
RANSAC-based point cloud segmentation. The latter is also used together with other point cloud features such as
eigenvalue based omnivariance, Mallet et al., 2011 to differentiate the remaining non-ground points into buildings and
vegetation. For the accuracy assessment of the automatic classification, two test areas, each 600 x 600 m, were selected
and manually revised. In a second semi-automatic step, the ground class is further subclassified by introducing three
additional classes: i ‘sealed ground’ e.g. roads, ii ‘water surface’, and iii ‘water body bottom’. For further analysis of
small training samples the heterogeneous classes of ‘sealed’ and ‘unsealed ground’ were semi-automatically classified into
‘green grass’, ‘dried up grass’, ’sand bare soil’, ’wetlands’, ‘darker asphalt’ and ‘lighter asphalt’.
For each channel, histograms showing the intensity value distribution for the seven classes are calculated and analyzed in
order to evaluate the potential of extending the geometrical classification approaches currently available for airborne
LiDAR mapping.
Because of their heterogeneity, the ground classes unsealed and sealed are examined in greater detail. Therefore, training areas
composed of different materials are manually selected and again histograms are calculated for each sub-class. The channel peak
values of each sub-class are then used for a supervised classification by pattern matching based on the Mahalanobis
distance.